The performance of GPU still slow even by keras fit_generator method

I have a large dataset 5GB that I want to use for training a neural network model designed using Keras. Although I am using Nvidia Tesla P100 GPU, the training is really slow (each epoch takes ~ 60-70s) (I choose the batch size=10000). After reading and searching, I found out that I can improve the training speed by using keras fit_generator instead of the typical fit. To do so, I coded the following:

from __future__ import print_function
import numpy as np
from keras import Sequential
from keras.layers import Dense
import keras
from sklearn.model_selection import train_test_split


def generator(C, r, batch_size):
    samples_per_epoch = C.shape[0]
    number_of_batches = samples_per_epoch / batch_size
    counter = 0

    while 1:
        X_batch = np.array(C[batch_size * counter:batch_size * (counter + 1)])
        y_batch = np.array(r[batch_size * counter:batch_size * (counter + 1)])
        counter += 1
        yield X_batch, y_batch

        # restart counter to yeild data in the next epoch as well
        if counter >= number_of_batches:
            counter = 0


if __name__ == "__main__":
    X, y = readDatasetFromFile()
    X_tr, X_ts, y_tr, y_ts = train_test_split(X, y, test_size=.2)

    model = Sequential()
    model.add(Dense(16, input_dim=X.shape[1]))
    model.add(keras.layers.advanced_activations.PReLU())
    model.add(Dense(16))
    model.add(keras.layers.advanced_activations.PReLU())
    model.add(Dense(16))
    model.add(keras.layers.advanced_activations.PReLU())
    model.add(Dense(1, activation='sigmoid'))

    model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])

    batch_size = 1000
    model.fit_generator(generator(X_tr, y_tr, batch_size), epochs=200, steps_per_epoch=X.shape[0]/ batch_size,
                        validation_data=generator(X_ts, y_ts, batch_size * 2),
                        validation_steps=X.shape[0] / batch_size * 2, verbose=2, use_multiprocessing=True)

    loss, accuracy = model.evaluate(X_ts, y_ts, verbose=0)
    print(loss, accuracy)

After running with fit_generator, the training time improved a little bit but it is still slow (each epoch now takes ~ 40-50s). When running nvidia-smi in the terminal, I found out that GPU utilization is ~15% only which makes me wonder if my code is wrong. I am posting my code above to kindly ask you if there is a bug causing to slow the performance of GPU.

Thank you,